RelCon: Relative Contrastive Learning for a Motion Foundation Model for Wearable Data

Summary

RelCon (Apple / UIUC / MIT, ICLR 2025) introduces Relative Contrastive Learning for training a motion FM from wrist-worn accelerometry. Rather than binary positive/negative contrastive pairs, RelCon trains a learnable semantic distance measure that captures motif similarity and domain-specific properties (rotation invariance), then uses this distance to define a continuous "relative" similarity between pairs.